feat(cloud-agent): 上下文计量 - 每步发出 context_pressure 读数并用上游实测定锚 (#601)
This commit is contained in:
AetherNo2332 authored and ddcat666 committed
1 parent
7ca625730e
commit
abf593c6c9
12 files changed
+1366
-15
No files matched your search
@@ -0,0 +1,296 @@
|
||||
package app
|
||||
|
||||
import (
|
||||
"encoding/json"
|
||||
"math"
|
||||
"unicode"
|
||||
"unicode/utf8"
|
||||
)
|
||||
|
||||
// cloudAgentContextPressure 是"这次请求还能装多少输入"的读数。
|
||||
// 消费方（上下文计量弹窗、运行诊断）直接读这份结构，不自己再算一遍。
|
||||
type cloudAgentContextPressure struct {
|
||||
EstimatedInputTokens int `json:"estimatedInputTokens"`
|
||||
ContextWindowTokens int `json:"contextWindowTokens"`
|
||||
ReservedOutputTokens int `json:"reservedOutputTokens"`
|
||||
UsableInputTokens int `json:"usableInputTokens"`
|
||||
PressureRatio float64 `json:"pressureRatio"`
|
||||
SourceBytes int `json:"sourceBytes"`
|
||||
PromptChars int `json:"promptChars"`
|
||||
PromptLimitChars int `json:"promptLimitChars"`
|
||||
ModelLimitConfigured bool `json:"modelLimitConfigured"`
|
||||
Estimate bool `json:"estimate"`
|
||||
// 三个窗口口径字段说明"这条读数按哪个算式算出来"：OverheadTokens 是工具 schema /
|
||||
// 协议包装 / 兜底轮的比例预留，InputBudgetTokens 才是真正可供输入使用的额度。
|
||||
OverheadTokens int `json:"overheadTokens,omitempty"`
|
||||
InputBudgetTokens int `json:"inputBudgetTokens,omitempty"`
|
||||
BudgetSource string `json:"budgetSource,omitempty"`
|
||||
// CompactAtTokens 是上游既有的"可重读正文就地裁剪"触发线（输入预算的 85%，见
|
||||
// compactCloudAgentContext）。它只是给读数一个参照刻度，不是本读数触发的动作。
|
||||
CompactAtTokens int `json:"compactAtTokens,omitempty"`
|
||||
}
|
||||
|
||||
// cloudAgentProjectedInputTokens 合成"下一步输入 token"的权威读数：
|
||||
// provider 锚点（模型自己的分词器读数）+ 本地估算的有符号增量；没有可用锚点时退回纯估算。
|
||||
// 增量带符号是刻意的：上下文被裁剪（图片移出、正文卸载）后本地估算会变小，
|
||||
// 投影也必须跟着变小，否则读数只会单调上涨。
|
||||
func cloudAgentProjectedInputTokens(pressure cloudAgentContextPressure, state *cloudAgentRuntime) (int, string) {
|
||||
if state == nil || state.TokenAnchor == nil || !state.TokenAnchor.Accepted {
|
||||
return pressure.EstimatedInputTokens, "estimate"
|
||||
}
|
||||
anchor := state.TokenAnchor
|
||||
delta := pressure.EstimatedInputTokens - anchor.EstimatedTokens
|
||||
return max(0, int(anchor.InputTokens)+delta), "provider"
|
||||
}
|
||||
|
||||
// cloudAgentContextPressurePayload 把读数摊成事件载荷。
|
||||
//
|
||||
// 两条硬规则（消费方不必猜）：
|
||||
// 1. 两个量纲各自命名、互不覆盖：本地估算描述**下一次**请求（readingScope=next_request /
|
||||
// estimateMethod=local_v1），上游实测量的是**上一次**请求（providerMeasurementScope=
|
||||
// previous_request），两者不能画成同一条曲线。
|
||||
// 2. 没有解析到模型自己声明的窗口时不下发窗口字段与占用率：宁可不给百分比，
|
||||
// 也不要拿兜底默认窗口编一个看起来很像真的"已占用 80%"。
|
||||
func cloudAgentContextPressurePayload(pressure cloudAgentContextPressure, state *cloudAgentRuntime, actual ...canonicalAgentRequest) map[string]any {
|
||||
payload := map[string]any{
|
||||
"estimatedInputTokens": pressure.EstimatedInputTokens,
|
||||
"sourceBytes": pressure.SourceBytes,
|
||||
"promptChars": pressure.PromptChars,
|
||||
"estimate": pressure.Estimate,
|
||||
// provider usage measures the previous request; the estimate describes
|
||||
// the next request. Consumers must not draw them as one series.
|
||||
"readingScope": "next_request", "estimateMethod": "local_v1",
|
||||
// 快照版本与相位：读数属于"下一次请求发出之前"的测量。
|
||||
"schemaVersion": 2, "phase": "before_request",
|
||||
}
|
||||
if pressure.PromptLimitChars > 0 {
|
||||
payload["promptLimitChars"] = pressure.PromptLimitChars
|
||||
}
|
||||
payload["modelLimitConfigured"] = pressure.ModelLimitConfigured
|
||||
if pressure.ModelLimitConfigured {
|
||||
payload["contextWindowTokens"] = pressure.ContextWindowTokens
|
||||
payload["reservedOutputTokens"] = pressure.ReservedOutputTokens
|
||||
payload["usableInputTokens"] = pressure.UsableInputTokens
|
||||
payload["inputBudgetTokens"] = pressure.InputBudgetTokens
|
||||
payload["overheadTokens"] = pressure.OverheadTokens
|
||||
payload["compactAtTokens"] = pressure.CompactAtTokens
|
||||
payload["budgetSource"] = firstNonEmpty(pressure.BudgetSource, "channel-model")
|
||||
if pressure.UsableInputTokens > 0 {
|
||||
payload["pressureRatio"] = pressure.PressureRatio
|
||||
}
|
||||
}
|
||||
if state == nil {
|
||||
return payload
|
||||
}
|
||||
// 上游实测锚点：provider 用模型自己的分词器报出的 prompt 规模，是权威读数。
|
||||
projectedTokens, tokenSource := cloudAgentProjectedInputTokens(pressure, state)
|
||||
if anchor := state.TokenAnchor; anchor != nil {
|
||||
payload["anchorStep"] = anchor.Step
|
||||
payload["anchor"] = map[string]any{
|
||||
"id": anchor.TaskID, "valid": anchor.Accepted, "ageSteps": max(0, state.Step-anchor.Step),
|
||||
}
|
||||
if anchor.Accepted {
|
||||
payload["pressureTokens"] = anchor.InputTokens
|
||||
payload["providerMeasurementScope"] = "previous_request"
|
||||
payload["providerUsage"] = map[string]any{
|
||||
"inputTokens": anchor.InputTokens, "cacheReadTokens": anchor.CachedTokens,
|
||||
"uncachedInputTokens": max(0, anchor.InputTokens-anchor.CachedTokens),
|
||||
"outputTokens": anchor.OutputTokens,
|
||||
}
|
||||
// anchorDeltaTokens 是"本地估算相对锚点那一步的增量"，与投影的分母无关：
|
||||
// 消费方拿它解释"较锚点 +N"，不能改成"投影 − 估算"。
|
||||
payload["anchorDeltaTokens"] = pressure.EstimatedInputTokens - anchor.EstimatedTokens
|
||||
if anchor.EstimatedTokens > 0 {
|
||||
payload["tokenScale"] = math.Round(float64(anchor.InputTokens)/float64(anchor.EstimatedTokens)*10000) / 10000
|
||||
}
|
||||
tokenSource = "provider"
|
||||
} else if anchor.RejectReason != "" {
|
||||
payload["anchorRejected"] = anchor.RejectReason
|
||||
}
|
||||
}
|
||||
payload["projectedTokens"] = projectedTokens
|
||||
payload["tokenSource"] = tokenSource
|
||||
// 两种量纲各自命名，消费方不需要靠 tokenSource 猜哪个数字属于哪把尺子。
|
||||
payload["measurementSource"] = tokenSource
|
||||
payload["normalizedInputTokens"] = pressure.EstimatedInputTokens
|
||||
if tokenSource == "provider" {
|
||||
payload["normalizedInputTokens"] = state.TokenAnchor.InputTokens
|
||||
}
|
||||
payload["projectedNextInputTokens"] = projectedTokens
|
||||
if pressure.UsableInputTokens > 0 && tokenSource == "provider" {
|
||||
payload["projectedPressureRatio"] = math.Min(9.99, float64(projectedTokens)/float64(pressure.UsableInputTokens))
|
||||
}
|
||||
payload["breakdown"] = cloudAgentContextBreakdownPayload(state, actual...)
|
||||
return payload
|
||||
}
|
||||
|
||||
// cloudAgentContextBreakdownPayload 报"这次要发出去的请求被谁占着"：三个 canonical 桶
|
||||
// 加上编译进系统提示的各分段，用的是与占用读数同一把估算尺子。
|
||||
// 桶合计与 canonical 总量的差是协议外壳（tool choice、缓存键），单独报而不摊进任何一个桶。
|
||||
func cloudAgentContextBreakdownPayload(state *cloudAgentRuntime, actual ...canonicalAgentRequest) map[string]any {
|
||||
canonical := state.Canonical
|
||||
if len(actual) > 0 {
|
||||
canonical = actual[0]
|
||||
}
|
||||
system := []byte(canonical.SystemPrompt)
|
||||
tools, _ := json.Marshal(canonical.Tools)
|
||||
messages, _ := json.Marshal(canonical.Messages)
|
||||
whole, _ := json.Marshal(canonical)
|
||||
systemTokens, toolTokens, messageTokens := estimateCloudAgentTokens(system), estimateCloudAgentTokens(tools), estimateCloudAgentTokens(messages)
|
||||
// 构成永远由本地估算给出；已有上游实测锚点时再给一份按锚点比例换算的读数，
|
||||
// 让展示口径与 provider 的计数同尺度（原值不覆盖）。
|
||||
scale := 1.0
|
||||
if anchor := state.TokenAnchor; anchor != nil && anchor.Accepted && anchor.EstimatedTokens > 0 {
|
||||
scale = float64(anchor.InputTokens) / float64(anchor.EstimatedTokens)
|
||||
}
|
||||
scaled := func(tokens int) int { return int(math.Round(float64(tokens) * scale)) }
|
||||
var systemSegments []cloudAgentContextSegment
|
||||
if segments := state.Policy.SystemSegments; len(segments) > 0 {
|
||||
// 分段必须把 system 桶填满：编译之外拼接的块（个人记忆）由
|
||||
// cloudAgentRecordMemorySegment 登记，剩余零头（标题、分隔与拼接文本）归入"其它"。
|
||||
// 否则分段合计会小于 system 桶，看起来像少算了一截。
|
||||
systemSegments = append([]cloudAgentContextSegment{}, segments...)
|
||||
accountedBytes, accountedTokens := 0, 0
|
||||
for _, segment := range systemSegments {
|
||||
accountedBytes += segment.Bytes
|
||||
accountedTokens += segment.Tokens
|
||||
}
|
||||
if remainder := len(system) - accountedBytes; remainder > 0 {
|
||||
systemSegments = append(systemSegments, cloudAgentContextSegment{
|
||||
Key: "other", Label: "其它（标题与拼接）", Bytes: remainder,
|
||||
Tokens: max(0, systemTokens-accountedTokens),
|
||||
})
|
||||
}
|
||||
// 逐段向上取整不满足可加性（Σ⌈ascii/4⌉ ≥ ⌈Σascii/4⌉），所以 system 桶的 token
|
||||
// 口径直接取分段合计：桶与分段是同一份分解，消费方会拿分段算占比，
|
||||
// 两边差几 token 就会显示成"分段占了 101%"。
|
||||
systemTokens = 0
|
||||
for index := range systemSegments {
|
||||
systemSegments[index].ScaledTokens = scaled(systemSegments[index].Tokens)
|
||||
systemTokens += systemSegments[index].Tokens
|
||||
}
|
||||
}
|
||||
buckets := []map[string]any{
|
||||
{"key": "system", "label": "系统提示（含画布摘要）", "bytes": len(system), "tokens": systemTokens, "scaledTokens": scaled(systemTokens)},
|
||||
{"key": "tools", "label": "工具 schema", "bytes": len(tools), "tokens": toolTokens, "scaledTokens": scaled(toolTokens)},
|
||||
{"key": "messages", "label": "会话消息（含工具结果）", "bytes": len(messages), "tokens": messageTokens, "scaledTokens": scaled(messageTokens)},
|
||||
}
|
||||
bucketBytes := len(system) + len(tools) + len(messages)
|
||||
breakdown := map[string]any{
|
||||
"totalBytes": len(whole), "totalTokens": estimateCloudAgentTokens(whole),
|
||||
"bucketBytes": bucketBytes, "bucketTokens": systemTokens + toolTokens + messageTokens,
|
||||
"envelopeBytes": max(0, len(whole)-bucketBytes),
|
||||
"buckets": buckets,
|
||||
"tokenScale": math.Round(scale*10000) / 10000,
|
||||
"scaledTotalTokens": scaled(estimateCloudAgentTokens(whole)),
|
||||
}
|
||||
if len(systemSegments) > 0 {
|
||||
breakdown["systemSegments"] = systemSegments
|
||||
}
|
||||
return breakdown
|
||||
}
|
||||
|
||||
// estimateCloudAgentTokens 是计量用的本地估算：ASCII 按 4 字节 1 token，CJK 等非 ASCII
|
||||
// 按 1 字符 1 token 计，数值上刻意保守（CJK 创作内容偏多）。
|
||||
//
|
||||
// 它与同文件的 cloudAgentEstimatedTokens 不是同一把尺子，也不要合并：后者是
|
||||
// **请求准入**的判据（非 ASCII 按 1.5 token/字符，宁可高估也不能把装不下的请求放出去），
|
||||
// 这里是**读数与锚点**的基准——锚点的采信区间（0.5×–2×）与 tokenScale 都是照着这把尺子
|
||||
// 标定的，换尺子等于把标定作废。它只是压力展示，既不是计费 token 数，
|
||||
// 也不能替代上游 tokenizer。
|
||||
func estimateCloudAgentTokens(value []byte) int {
|
||||
if len(value) == 0 {
|
||||
return 0
|
||||
}
|
||||
asciiUnits, nonASCII := 0, 0
|
||||
for len(value) > 0 {
|
||||
r, size := utf8.DecodeRune(value)
|
||||
if r == utf8.RuneError && size == 1 {
|
||||
asciiUnits++
|
||||
value = value[1:]
|
||||
continue
|
||||
}
|
||||
value = value[size:]
|
||||
if r <= unicode.MaxASCII {
|
||||
asciiUnits++
|
||||
} else {
|
||||
nonASCII++
|
||||
}
|
||||
}
|
||||
return int(math.Ceil(float64(asciiUnits)/4.0)) + nonASCII
|
||||
}
|
||||
|
||||
// cloudAgentContextPressure 解析"这次请求能用多少输入"。
|
||||
//
|
||||
// 没有解析到真实窗口时 UsableInputTokens 保持 0 且 ModelLimitConfigured=false，
|
||||
// 消费方据此显示"未配置模型上限"而不是编一个占用率。
|
||||
func (s *Service) cloudAgentContextPressure(canonical canonicalAgentRequest, prompt string, request CloudAgentRequest) cloudAgentContextPressure {
|
||||
raw, _ := json.Marshal(canonical)
|
||||
pressure := cloudAgentContextPressure{
|
||||
EstimatedInputTokens: estimateCloudAgentTokens(raw),
|
||||
SourceBytes: len(raw),
|
||||
PromptChars: utf8.RuneCountInString(prompt),
|
||||
Estimate: true,
|
||||
}
|
||||
// 提示词字符上限只来自渠道模型能力（逻辑模型没有单一渠道模型，保持 0 = 不声明）。
|
||||
if text := s.cloudAgentChannelTextCapability(request); text != nil {
|
||||
pressure.PromptLimitChars = text.References.PromptMaxChars
|
||||
}
|
||||
budget := s.cloudAgentContextBudgetForRequest(request)
|
||||
if !budget.Configured {
|
||||
return pressure
|
||||
}
|
||||
pressure.ModelLimitConfigured = true
|
||||
pressure.ContextWindowTokens = budget.ContextWindowTokens
|
||||
pressure.ReservedOutputTokens = budget.MaxOutputTokens
|
||||
pressure.OverheadTokens = budget.OverheadTokens
|
||||
pressure.InputBudgetTokens = budget.InputBudgetTokens
|
||||
pressure.CompactAtTokens = budget.CompactAtTokens
|
||||
pressure.BudgetSource = budget.Source
|
||||
// 展示口径与请求准入判据共用同一份输入预算：否则会出现"界面 79%、后台按 82% 拒绝"。
|
||||
pressure.UsableInputTokens = budget.InputBudgetTokens
|
||||
if pressure.UsableInputTokens > 0 {
|
||||
pressure.PressureRatio = math.Min(9.99, float64(pressure.EstimatedInputTokens)/float64(pressure.UsableInputTokens))
|
||||
}
|
||||
return pressure
|
||||
}
|
||||
|
||||
// canonicalAgentRequestFromInput 从已落库的任务输入里取回"实际发出去的那份 canonical"：
|
||||
// 压力读数与锚点都必须对着真实信封，而不是内存里另算一份可能与准入改写结果不同的副本。
|
||||
func canonicalAgentRequestFromInput(input map[string]any) (canonicalAgentRequest, bool) {
|
||||
requests, ok := input["agentRequests"].(map[string]any)
|
||||
if !ok {
|
||||
return canonicalAgentRequest{}, false
|
||||
}
|
||||
if requests["canonical"] == nil {
|
||||
return canonicalAgentRequest{}, false
|
||||
}
|
||||
raw, err := json.Marshal(requests["canonical"])
|
||||
if err != nil {
|
||||
return canonicalAgentRequest{}, false
|
||||
}
|
||||
var canonical canonicalAgentRequest
|
||||
if json.Unmarshal(raw, &canonical) != nil {
|
||||
return canonical, false
|
||||
}
|
||||
return canonical, true
|
||||
}
|
||||
|
||||
// cloudAgentNoteContextWindowResolved 在"窗口从未确认变为已确认"时落一条 context_transition。
|
||||
//
|
||||
// 这一步只在读数口径变化时发生（例如早先拿不到模型能力、后来解析到了窗口）：真实上下文
|
||||
// 没有变小，消费方必须能把它标成"模型窗口已识别"，否则一识别就被画成上下文骤降。
|
||||
func cloudAgentNoteContextWindowResolved(runID string, state *cloudAgentRuntime, pressure cloudAgentContextPressure) {
|
||||
if state == nil || runID == "" || !pressure.ModelLimitConfigured || state.ContextWindowKnown {
|
||||
return
|
||||
}
|
||||
state.ContextWindowKnown = true
|
||||
state.event(runID, "context_transition", map[string]any{
|
||||
"kind": "window_resolved", "reason": "window_resolved",
|
||||
"after": map[string]any{
|
||||
"contextWindowTokens": pressure.ContextWindowTokens, "usableInputTokens": pressure.UsableInputTokens,
|
||||
"source": firstNonEmpty(pressure.BudgetSource, "channel-model"),
|
||||
},
|
||||
"text": "模型窗口已识别：读数改用窗口口径，这不是上下文变少",
|
||||
})
|
||||
}
|
||||
Reference in new issue
Block a user